cutile-forge
Tile-Fusion Graph compiler + hybrid PCIe-streaming LLM engine on NVIDIA CuTile (sm_89 ↔ sm_120). 8 CUDA kernels → 3 Python tiles; 117 tok/s 7B, 70B in one 32 GB GPU at 80.8% of PCIe 5.0 ceiling; torch.compile backend.
- 0
- Python
- MIT
- 2 d ago
